Marketing scenario-oriented ai agent construction method, system, device and medium

Abstract

The application discloses a marketing-scene-oriented AI intelligent agent construction method, and belongs to the technical field of data processing, which comprises the following steps: performing desensitization processing on multi-source original data to generate desensitized behavior data set, desensitized transaction data set and desensitized device environment metadata; calculating behavior mode similarity based on the desensitized behavior data set and the desensitized device environment metadata, constructing a cross-device user identification association graph, and generating a global user identification; generating a dynamic user portrait encapsulating a vector similarity matching value for the global user identification; inputting the vector similarity matching value and the business target priority into a constraint satisfaction solver to obtain a strategy parameter; inputting the strategy parameter into a generative large model to output a marketing scheme; collecting positive feedback data and negative feedback data of users on marketing intervention, and selecting an updated strategy. The application can improve the accuracy of the marketing strategy. The application also discloses a system, an electronic device and a computer readable storage medium for implementing the above method.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system, device and medium for constructing AI intelligent agents for marketing scenarios. Background Technology

[0002] An AI agent is an artificial intelligence system capable of self-perceiving its environment, making decisions, and executing tasks. It can not only understand instructions and passively answer questions, but also think independently, utilize tools, and proactively solve problems like a human. It typically consists of modules for perception, planning, memory, and tool usage, enabling automated processing of complex tasks such as cross-application operations and data analysis. Its core characteristics are proactive thinking and cross-tool operation capabilities, distinguishing it from traditional AI systems that rely on pre-set instructions. Therefore, applying AI agents in the marketing field allows for autonomous goal setting, high-value user screening, and personalized targeting strategies. This reduces human intervention while achieving automated and precise reach, thereby improving conversion rates.

[0003] However, current AI-based marketing decision-making systems are not perfect and still have significant shortcomings when facing complex scenarios with multiple marketing touchpoints (such as apps, mini-programs, H5, and web) and across devices. For example, existing solutions cannot effectively integrate multi-source data from various marketing touchpoints (such as user behavior data streams, transaction data streams, and device environment streams), and lack data anonymization processing, posing a risk of privacy leaks. Traditional AI agents mostly rely on static rules or single models and lack the ability to associate user identifiers across devices, resulting in fragmented tags for the same user on different devices, affecting the accuracy of marketing strategies. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a method, system, device, and medium for constructing AI intelligent agents for marketing scenarios. The technical problem to be solved by this invention is achieved through the following technical solution: The first aspect of this invention provides a method for constructing an AI agent for marketing scenarios, comprising: The collected raw data from multiple marketing touchpoints is anonymized to generate an anonymized dataset; wherein, the raw data from multiple sources includes user behavior data stream, user transaction data stream, and device environment data stream; the anonymized dataset includes anonymized behavior dataset, anonymized transaction dataset, and anonymized device environment metadata; The behavior pattern similarity is calculated based on the de-identified behavior dataset and the de-identified device environment metadata. A cross-device user identifier association graph is constructed based on the behavior pattern similarity, multiple global user identifiers are generated, and the multiple global user identifiers are used as association indexes and bound to the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata respectively. Using the global user identifier as the primary key, a corresponding dynamic user profile is generated for each global user identifier; wherein, the dynamic user profile encapsulates a vector similarity matching value; The strategy parameters are obtained by solving the solution using the vector similarity matching values ​​and the preset business objective priority input constraints; The strategy parameters are encoded into semi-structured prompt text, the semi-structured prompt text is input into a generative large model, and a marketing plan is output so that the multi-channel reach gateway can execute the marketing plan to reach users. Collect positive and negative feedback data from users regarding marketing interventions, calculate the feedback density, and adaptively select an update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver; wherein, the feedback density is the ratio of the sum of the number of times users generate positive and negative feedback within a preset first sliding time window to the total number of marketing interventions issued during the same period.

[0005] The method provided by this invention introduces the concept of AI agent construction, realizing a full-link intelligent upgrade from data collection to strategy generation. By anonymizing multi-source raw data, user privacy and security are ensured, while providing a high-quality, global data foundation for the AI ​​agent. By calculating the similarity of behavioral patterns using anonymized behavioral data and anonymized device environment metadata, a cross-device association graph is constructed, solving the problem of multi-terminal data silos and enabling the AI ​​agent to make accurate decisions based on global user identifiers. Multi-objective optimization is performed using a constraint satisfaction solver (CSP), ensuring the compliance and interpretability of marketing strategies. An adaptive update mechanism for strategy parameters based on feedback density is introduced, improving the robustness of the model.

[0006] In one possible implementation, the process of de-identifying the multi-source raw data collected from various marketing touchpoints to generate a de-identified dataset includes: Identify the data to be classified in the user behavior data, and select a de-identification strategy for the data to be classified based on the data distribution characteristics; wherein, the data to be classified includes direct identification information, indirect identification information, and behavior frequency statistics. If the data to be classified is the direct identification information, then the first desensitization strategy is selected, and the first desensitization strategy is to perform irreversible hash replacement or encrypted erasure processing on the direct identification information; If the data to be classified is the indirect identification information, then the second desensitization strategy is selected. The second desensitization strategy includes automatically binning continuous values ​​into discrete intervals according to data distribution, truncating timestamps to hourly or daily levels according to business scenarios, and generalizing geographical locations to city or provincial levels according to population density. If the data to be classified is the behavior frequency statistics, then the third desensitization strategy is selected, which is to inject differential privacy noise into the behavior frequency statistics.

[0007] In one possible implementation, the step of calculating behavioral pattern similarity based on the de-identified behavioral dataset, constructing a cross-device user identifier association graph based on the behavioral pattern similarity, and generating multiple global user identifiers includes: Multiple records without assigned global user identifiers are read from the de-identified behavior dataset. Behavioral metadata for each record is extracted and aggregated at the granularity of session identifier or device fingerprint hash to obtain a behavior pattern vector. The behavioral metadata includes the session identifier, the identifier of the interacted object, the type of behavior action, and the binned page dwell time. Calculate the behavioral pattern similarity between any two behavioral pattern vectors; Each session identifier or device fingerprint hash is used as a node in the cross-device user identifier association graph. An undirected edge is added to connect two nodes whose behavior pattern similarity is higher than a preset similarity threshold to obtain the cross-device user identifier association graph. Perform connected component analysis or community discovery algorithm on the cross-device user identifier association graph, merge all nodes in the same connected component or the same community into the same logical user, and assign a unique global user identifier to each logical user. At the same time, write the global user identifier back to the index field of the corresponding record in the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata. For any node in the cross-device user identifier association graph that has not been assigned a global user identifier, obtain the set of candidate global user identifiers for all nodes connected to that node, calculate the initial attribution probability of that node for each candidate global user identifier based on the edge weights, and perform attenuation correction on each of the initial attribution probabilities based on the shared device factor to obtain the corrected attribution probability; wherein, the shared device factor is the number of different global user identifiers bound to the device fingerprint hash of the physical terminal to which the node belongs; the edge weight is the behavioral pattern similarity between the node connected to it and the node; If the candidate global user identifier set is empty or the maximum value of the corrected attribution probability is lower than the preset merging threshold, a new global user identifier is assigned to the node; otherwise, the candidate global user identifier corresponding to the maximum value of the corrected attribution probability is used as the global user identifier of the node.

[0008] In one possible implementation, the dynamic user profile further encapsulates core behavioral representation fields, lifecycle stage fields, recent positive feedback response rate, recent negative feedback response rate, and marketing fatigue fields; the step of generating a corresponding dynamic user profile for each global user identifier as the primary key includes: Long-term attribute field, short-term interest field and real-time intent field are obtained respectively; the long-term behavior representation vector in the long-term attribute field, the short-term interest vector in the short-term interest field and the real-time behavior embedding vector in the real-time intent field are encoded and normalized respectively; the normalized vectors are weighted and fused according to preset weight coefficients to generate a fixed-dimensional behavior pattern vector as the core behavior representation field. Historical transaction data of the target user is read from the anonymized transaction dataset. Based on the historical transaction data and preset time window rules, the customer relationship stage of the target user is determined and the customer relationship stage is marked as the lifecycle stage field. The historical transaction data includes registration time, first transaction time, last transaction time and cumulative number of transactions. The similarity between the core behavior representation field and the preset typical user group center vector is calculated to obtain the vector similarity matching value; The proportion of positive feedback generated by users within a preset second sliding time window to the total number of effective marketing campaigns delivered during the same period is calculated as the recent positive feedback response rate. The proportion of the number of times users generate negative feedback within a preset third sliding time window to the total number of effective marketing campaigns delivered during the same period is calculated as the recent negative feedback response rate. The marketing fatigue level is calculated based on whether the user exhibits negative feedback behavior within the preset fourth sliding time window, the preset upper limit for marketing intervention per unit time, and the total number of effective marketing campaigns delivered during the same period.

[0009] In one possible implementation, the real-time intent field is obtained as follows: Configure a lightweight sequence model to extract the behavior sequence of the current session from the de-identified behavior dataset, encode it to generate a real-time behavior embedding vector as the real-time intent field; and / or; The short-term interest field is obtained using the following method: Configure a nearline computing engine, call the time decay function to extract records from the de-identified behavior dataset that match the global user identifier within the most recent preset time period; group and aggregate the matching records based on category identifier or inventory unit identifier, and assign different behavior weights to click behavior and add-to-cart behavior according to behavior type, and then perform weighted summation to obtain the interest score for each category or inventory unit; correct the interest score based on the page dwell time after binning; generate a short-term interest vector as the short-term interest field based on the corrected interest score for each category or inventory unit; and / or; The long-term attribute field is obtained using the following method: Configure an offline computing engine, call the value computing model, and use the core behavior representation field and the historical transaction aggregation features in the anonymized transaction dataset that match the global user identifier as input data. Output the value stratification label and the life cycle stage label as the long-term attribute field; wherein, the historical transaction aggregation features include consumption frequency, consumption amount and recent purchase interval.

[0010] In one possible implementation, the adaptive selection and update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver includes: If the feedback density is higher than the first preset threshold, then online reinforcement learning is used to update the parameters of the constraint satisfaction solver. If the feedback density is not higher than the first preset threshold and not lower than the second preset threshold, then the user vector similarity is calculated based on the core behavior representation field encapsulated by the dynamic user profile to perform similar user clustering, and counterfactual strategy exploration is performed to update the parameters of the constraint satisfaction solver; wherein, the first preset threshold is greater than the second preset threshold; If the feedback density is lower than the second preset threshold, an offline retraining process is triggered to reconstruct the offline teacher model, and the knowledge of the offline teacher model is transferred to the online lightweight student model through knowledge distillation; wherein, the online lightweight student model is the online running version of the constraint satisfaction solver.

[0011] In one possible implementation, the knowledge transfer from the offline teacher model to the online lightweight student model via knowledge distillation includes: The parameters of the offline teacher model are frozen. Taking the user state sampling set as input, the parameters of the online lightweight student model are updated with the goal of minimizing the difference between the output distribution of the online lightweight student model and the output distribution of the offline teacher model. Each user state vector in the user state sampling set is obtained by concatenating or encoding the core behavior representation field, vector similarity matching value, life cycle stage field, recent positive feedback response rate, recent negative feedback response rate, and marketing fatigue field encapsulated by the dynamic user profile of the corresponding user.

[0012] A second aspect of this invention provides an AI agent construction system for marketing scenarios, comprising: The multi-source data anonymization module is used to anonymize the raw data collected from various marketing touchpoints from multiple sources to generate an anonymized dataset; wherein, the raw data from multiple sources includes user behavior data streams, user transaction data streams, and device environment streams; the anonymized dataset includes an anonymized behavior dataset, an anonymized transaction dataset, and anonymized device environment metadata; The global user identifier generation module is used to calculate the similarity of behavior patterns based on the de-identified behavior dataset and the de-identified device environment metadata, construct a cross-device user identifier association graph based on the behavior pattern similarity, generate multiple global user identifiers, and bind the multiple global user identifiers as association indexes to the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata respectively; The dynamic user profile generation module is used to generate a corresponding dynamic user profile for each global user identifier, using the global user identifier as the primary key; wherein, the dynamic user profile encapsulates a vector similarity matching value; The constraint satisfaction solution module is used to input the vector similarity matching value and the preset business objective priority into the constraint satisfaction solver to obtain the strategy parameters; The marketing plan generation module is used to encode the strategy parameters into semi-structured prompt text, input the semi-structured prompt text into the generative big model, and output the marketing plan so that the multi-channel reach gateway can execute the reach of the marketing plan; The feedback parameter tuning module is used to collect positive and negative feedback data from users regarding marketing interventions, calculate the feedback density, and adaptively select an update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver. The feedback density is the ratio of the sum of the number of times users generate positive and negative feedback within a preset first sliding time window to the total number of marketing interventions issued during the same period.

[0013] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement an AI agent construction method for marketing scenarios provided in the first aspect of the present invention.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an AI agent construction method for marketing scenarios provided in the first aspect of the present invention.

[0015] For a detailed description of the second to fourth aspects of the present invention and their various implementations, please refer to the detailed description in the first aspect and its various implementations; and for a detailed description of the beneficial effects of the second to fourth aspects and their various implementations, please refer to the beneficial effect analysis in the first aspect and its various implementations, which will not be repeated here.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an AI agent construction method for marketing scenarios according to an embodiment of the present invention. Figure 2 This is a structural block diagram of an AI intelligent agent construction system for marketing scenarios according to an embodiment of the present invention; Figure 3 This is a block diagram of the internal structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0019] This invention provides a method for constructing an AI agent for marketing scenarios. This method is applied to electronic devices, which can be servers or terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal device can be a smartphone, tablet, desktop computer, wearable device, etc., but is not limited to these.

[0020] Figure 1 This is a flowchart illustrating a method for constructing an AI agent for marketing scenarios, as provided in this embodiment. Figure 1 As shown, the main process of this method is described below (steps S101 to S106): Step S101: De-identify the multi-source raw data collected from various marketing touchpoints to generate a de-identified dataset; wherein, the multi-source raw data includes user behavior data stream, user transaction data stream, and device environment stream; the de-identified dataset includes de-identified behavior dataset, de-identified transaction dataset, and de-identified device environment metadata; Step S102: Calculate the similarity of behavior patterns based on the de-identified behavior dataset and the de-identified device environment metadata; construct a cross-device user identifier association graph based on the behavior pattern similarity; generate multiple global user identifiers; and bind the multiple global user identifiers as association indexes to the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata, respectively. Step S103: Using the global user identifier as the primary key, generate a corresponding dynamic user profile for each global user identifier; wherein, the dynamic user profile encapsulates a vector similarity matching value; Step S104: The vector similarity matching value and the preset business target priority input constraints are satisfied by the solver to obtain the strategy parameters; Step S105: Encode the strategy parameters into semi-structured prompt text, input the semi-structured prompt text into the generative big model, and output the marketing plan so that the multi-channel reach gateway can execute the marketing plan to reach users. Step S106: Collect positive and negative feedback data from users regarding marketing interventions, calculate the feedback density, and adaptively select an update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver; wherein, the feedback density is the ratio of the sum of the number of times users generate positive and negative feedback within a preset first sliding time window to the total number of marketing interventions issued during the same period.

[0021] In this embodiment, the behavioral data stream refers to page browsing trajectory, click behavior sequence, duration of stay on each page, and records of adding to cart and favorites operations; the transaction data stream refers to transaction behavior and conversion behavior data such as order amount, payment time, purchase frequency, and order cancellation markers; the device environment stream refers to information collected along with the user's access request that describes the terminal device and network context, including at least device identifiers (such as MAC address, IMEI, advertising identifier), device model, operating system type and version, screen resolution, User-Agent string, IP address, approximate geographical location, and application version number device identifier.

[0022] In some optional embodiments, for step S101, the data to be classified in the user behavior data is first identified, and a desensitization strategy for the data to be classified is selected based on the data distribution characteristics; wherein, the data to be classified includes direct identification information, indirect identification information and behavior frequency statistics.

[0023] In this embodiment, direct identification information refers to information that can identify a specific natural person on its own, including but not limited to name, ID number, mobile phone number, email address, device MAC address, OpenID, etc.; indirect identification information refers to attributes that cannot identify a specific natural person on its own, but can be inferred through link attacks when combined with other records, including but not limited to precise age, date of birth, GPS coordinates, postal code, gender, occupation, registration time, etc.; behavior frequency statistics refer to the aggregated statistical value of the number of times or the cumulative amount of a certain type of user behavior occurs within a preset time window, such as the number of clicks, the number of times added to cart, the number of times reached, the cumulative value of dwell time, etc.

[0024] If the data to be classified is direct identification information, then the first desensitization strategy is selected. The first desensitization strategy is to perform irreversible hash replacement or encrypted erasure processing on the direct identification information to achieve de-identification.

[0025] If the data to be classified is indirect identification information, then the second desensitization strategy is selected. The second desensitization strategy includes automatically binning continuous values ​​into discrete intervals according to data distribution, truncating timestamps to hourly or daily levels according to business scenarios, and generalizing geographical locations to city or provincial levels according to population density.

[0026] If the data to be classified is behavioral frequency statistics, then the third desensitization strategy is selected. This strategy involves injecting differential privacy noise into the behavioral frequency statistics. Specifically, it superimposes random perturbation values ​​following a specific probability distribution (such as a Laplace or Gaussian distribution) onto the calculated results of behavioral frequency statistics, aggregated counts, or machine learning gradients, ensuring that the final published values ​​satisfy ε-differential privacy (ε...). DP) or (ε,δ)-differential privacy ((ε,δ) DP (Approximate Differential Privacy) is defined to prevent member inference attacks and further reduce the risk of re-identification.

[0027] In some optional embodiments, for step S102, multiple records that have not been assigned a global user identifier are read from the de-identified behavior dataset, behavior metadata of each record is extracted, and behavior pattern vectors are aggregated at the granularity of session identifier or device fingerprint hash; wherein, behavior metadata includes session identifier, the identifier of the interacted object, the type of behavior action, and the page dwell time after binning; the behavior pattern similarity between any two behavior pattern vectors is calculated; each session identifier or device fingerprint hash is used as a node in the cross-device user identifier association graph, and an undirected edge is added to connect two nodes whose behavior pattern similarity is higher than a preset similarity threshold to obtain the cross-device user identifier association graph.

[0028] Among them, the session identifier refers to the temporary unique identifier (session identifier) ​​assigned by the server or front-end tracking point to a user's continuous access interaction process. It is used to group page browsing, searching, adding to cart, claiming coupons and staying behavior generated within the same time period into the same behavior sequence.

[0029] The identifier of the interacted object refers to the identifier of the specific business object affected by the user's behavior. Typical identifiers include product SKU identifier (SKU_ID), third-level category identifier (Category_ID), search term identifier (Query_ID), or activity page identifier (Activity_ID), depending on the aggregation granularity.

[0030] Behavioral action type refers to the types of behavioral actions that users take at marketing touchpoints, including at least page views / clicks, adding to cart, and favorites.

[0031] Binning-based page dwell time refers to the location identifier or interval code obtained by discretizing the original dwell time (usually in milliseconds) of a user on a certain page or product details page according to a preset or automatically learned numerical range (binning boundary); for example, the original dwell time is mapped to one of four levels: 1-5s, 5-30s, 30-120s, and >120s. The binning boundary is automatically determined by the system in the offline stage based on the statistical distribution of the total dwell time (such as equal frequency quantiles or equal width intervals), and can be recalculated based on online data drift detection.

[0032] Optionally, for calculating the behavioral pattern similarity between any two behavioral pattern vectors, an interaction object set is constructed at the granularity of session identifier or device fingerprint hash, based on the identifiers of the interacted objects recorded at the same granularity. Behavior sequences are obtained by sorting the behavior action types by time, and a dwell time distribution vector is obtained by statistically analyzing the binned page dwell time. These together constitute the behavioral pattern vector. Then, based on the Jaccard similarity of the interaction object set, the edit distance similarity of the behavior sequence, and the cosine similarity or Euclidean distance similarity of the dwell time distribution vector, these three similarities are weighted, fused, and normalized to obtain a comprehensive behavioral pattern similarity score. Undirected edges are added to the cross-device user identifier association graph for node pairs with comprehensive behavioral pattern similarity scores higher than a preset similarity threshold.

[0033] When adding an undirected edge connecting two nodes in a cross-device user identifier association graph, the weight attribute of this edge is set to the value of behavioral pattern similarity. The edge weight value is between 0 and 1. The larger the weight, the closer the behavioral patterns represented by the two nodes are, and the higher the confidence level of classifying them as the same natural person. In subsequent connected component analysis or community detection (such as Louvain, Label Propagation), the community partitioning strength can be adjusted based on the edge weight or used as an input factor for calculating the attribution probability. Node pairs with values ​​below the threshold are not added to the edge and are considered to be not yet associated.

[0034] The cross-device user identification association graph is constructed based on the de-identified behavioral dataset because behavioral data streams are denser and more individual-discriminative in terms of time and device dimensions, making them suitable as behavioral fingerprints for similarity calculation. In contrast, transaction datasets are usually sparse and greatly affected by payment methods and account binding. If they are used as the main association basis, a large number of users who have not placed orders or new devices will not be correctly associated, reducing the system's coverage and accuracy.

[0035] Then, perform connectivity component analysis or community discovery algorithms on the cross-device user identifier association graph, merge all nodes in the same connectivity component or the same community into the same logical user, assign a unique global user identifier to each logical user, and write the global user identifier back to the index field of the corresponding record in the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata.

[0036] In this embodiment, the global user identifier refers to a logical user number that uniquely represents a natural person in the enterprise's marketing system, generated based on a cross-device user identifier association graph. This identifier is obtained by graph association and merging of the behavioral data of the same user under different terminals and different cookies / device IDs.

[0037] For any node in the cross-device user identifier association graph that has not been assigned a global user identifier, obtain the set of candidate global user identifiers for all nodes connected to that node. Calculate the initial attribution probability of the node for each candidate global user identifier based on the edge weights. Attenuate and correct each initial attribution probability based on the shared device factor to obtain the corrected attribution probability. Here, the shared device factor is the number of different global user identifiers bound to the device fingerprint hash of the physical terminal to which the node belongs; the edge weight is the behavioral pattern similarity between the node and the node connected to it. If the set of candidate global user identifiers is empty or the maximum value of the corrected attribution probability is lower than a preset merging threshold, a new global user identifier is assigned to the node; otherwise, the candidate global user identifier corresponding to the maximum value of the corrected attribution probability is used as the global user identifier of the node.

[0038] When a device fingerprint hash in the cross-device user identifier association graph is bound to at least two global user identifiers, the physical terminal is determined to be a shared device (such as a home tablet or a public display terminal). The attribution probability of each node on the device is then corrected. For example, the probability of attribution to any existing global user identifier is multiplied by a decay coefficient that monotonically decreases as the shared device factor (the number of different global user identifiers bound to the shared device) increases, or the prior probability of creating a new independent global user identifier is increased. This reduces the risk of different natural persons' sessions on the shared device being mistakenly merged into the same global user identifier. It can reduce the mixing of non-personal behavior data into dynamic user profiles, ensuring that subsequent recommendation strategies, value stratification, and lifecycle judgments are based on the correct natural person attribution. At the same time, when a user deletes or withdraws consent requests, the corresponding user data can be accurately cascaded and cleared according to the global user identifier without affecting other users sharing the same terminal.

[0039] In some optional embodiments, the dynamic user profile also encapsulates core behavioral representation fields, lifecycle stage fields, recent positive feedback response rate, recent negative feedback response rate, and marketing fatigue fields; the generation methods for each field are as follows: For the core behavior representation field, the long-term attribute field, short-term interest field, and real-time intent field are obtained respectively; the long-term behavior representation vector in the long-term attribute field, the short-term interest vector in the short-term interest field, and the real-time behavior embedding vector in the real-time intent field are encoded and normalized respectively; the normalized vectors are weighted and fused according to the preset weight coefficients to generate a fixed-dimensional behavior pattern vector as the core behavior representation field.

[0040] Optionally, configure a lightweight sequence model (LSTM, GRU, or shallow Transformer encoder) to extract the behavior sequence of the current session from the anonymized behavior dataset, and encode it to generate a real-time behavior embedding vector as a real-time intent field; configure a nearline computing engine to call a time decay function to extract records in the anonymized behavior dataset that match the global user identifier within the most recent preset time period; group and aggregate the matching records based on category identifier or inventory unit identifier, and assign different behavior weights to click behavior and add-to-cart behavior according to behavior type, and then perform weighted summation to obtain the interest score for each category or inventory unit; adjust the interest score based on the binned page dwell time; generate a short-term interest vector as a short-term interest field based on the adjusted interest score for each category or inventory unit; configure an offline computing engine to call a value calculation model (such as RFM model, deep value network, or survival analysis model with decay), using the core behavior representation field and the historical transaction aggregation features in the anonymized transaction dataset that match the global user identifier as input data, and outputting value stratification labels and lifecycle stage labels as long-term attribute fields; among which, the historical transaction aggregation features include consumption frequency, consumption amount, and recent purchase interval.

[0041] For the lifecycle stage field, the target user's historical transaction data is read from the anonymized transaction dataset. Based on the historical transaction data and the preset time window rules, the customer relationship stage of the target user is determined and marked as the lifecycle stage field. The historical transaction data includes the registration time, the first transaction time, the last transaction time, and the cumulative number of transactions.

[0042] The preset time window rules include at least the new customer determination threshold, the inactive customer determination threshold, and the churn determination threshold. If the cumulative number of transactions is zero and the time since registration has not exceeded the preset observation period, the customer is determined to be a potential user. If the cumulative number of transactions is ≥1 and the time since the last transaction does not exceed the new customer determination threshold, the customer is determined to be a new customer. If the cumulative number of transactions is ≥2 and the time since the last transaction does not exceed the inactive customer determination threshold, the customer is determined to be an active customer. If the cumulative number of transactions is ≥2 and falls between the inactive customer determination threshold and the churn determination threshold, the customer is determined to be an inactive customer. If the cumulative number of transactions is ≥2 and exceeds the churn determination threshold, the customer is determined to be a churned customer.

[0043] For vector similarity matching values, the similarity between the core behavior representation field and the preset typical user group center vector is calculated (e.g., cosine similarity or normalized Euclidean distance similarity, with a value range of [0,1]), and the vector similarity matching value is obtained.

[0044] Among them, the typical user cluster center vector refers to the mean or weighted center vector of all sample vectors in each cluster after performing cluster analysis (such as K-means, GMM or HDBSCAN) on the core behavioral representation fields of historical user samples (behavioral pattern vectors obtained by fusing long-term attribute fields, short-term interest fields and real-time intent fields). Each typical user cluster center vector is associated with a cluster identifier and a cluster semantic label (such as high-net-worth users aged 25-35 in first-tier cities with maternal and infant preferences).

[0045] The recent positive feedback response rate is calculated as the proportion of positive feedback generated by users within a preset second sliding time window to the total number of effective marketing campaigns delivered during the same period.

[0046] The recent negative feedback response rate is calculated as the proportion of the number of times users generate negative feedback within the preset third sliding time window to the total number of effective marketing campaigns delivered during the same period.

[0047] Marketing fatigue is calculated based on whether users exhibit negative feedback behavior within a preset fourth sliding time window, the preset upper limit for marketing intervention per unit time, and the total number of effective marketing campaigns delivered during the same period.

[0048] In this embodiment, marketing intervention refers to the complete marketing action that the system sends to the target user through the multi-channel reach gateway. The multi-channel reach gateway refers to the middleware / service layer that uniformly encapsulates the sending channels in the marketing system. It provides a unified calling interface to the outside world and internally completes channel routing, protocol adaptation, rate limiting and retries, and status receipt collection. It pushes the marketing content generated upstream to channels such as SMS, APP Push, WeChat (official account / mini program / service account), email, and Enterprise WeChat according to specified priorities / rules.

[0049] Positive feedback indicates that users have a positive attitude towards the marketing intervention or are motivated to perform the expected behavior. Examples include clicking links, buttons, or product cards in marketing messages; claiming coupons; adding items to cart; adding to favorites; placing orders; following or subscribing to the corresponding brand account (e.g., initial WeChat subscription authorization); and entering the product details page and staying there for more than a certain amount of time. Negative feedback indicates that users have a negative attitude towards the marketing intervention; they may feel aversion or the marketing has not produced the expected results. Examples include unsubscribing; marking marketing SMS messages as spam; explicitly selecting "not interested"; disabling push notifications; blocking the official account; messages being delivered but not opened (long-term zero open rate); multiple marketing campaigns without any clicks / conversions (which can accumulate as negative weight); uninstalling the app or unfollowing the official account after receiving the marketing (if there is evidence of event tracking).

[0050] It should be noted that the various sliding time windows in this embodiment can be set to be the same or different, and this embodiment does not make specific limitations.

[0051] In some optional embodiments, the constraint satisfaction solver models the marketing strategy generation as a constraint satisfaction problem, defining the following triples (X, D, C): The variable set X consists of marketing strategy parameters to be solved, including at least the user segment identifier (xseg), recommended product list (xprod), discount range (xdisc), channel priority vector (xchan), and brand tone parameter (xbrand). The user segment identifier is determined by the system based on the vector similarity matching value of the current user from several predefined typical user segments (such as high-value active customers, price-sensitive new customers, and churned customers), and is assigned a segment number or name. The channel priority vector refers to the system's configuration parameters for the channels allowed for a single user or user segment within the same marketing campaign cycle, including their order, weight, and whether they are enabled. The brand tone parameter refers to configuration parameters related to brand image / compliance requirements, such as the style of marketing copy, discount expression, visual elements, and discount minimum, ensuring that the marketing plan output by the generative large model aligns with the brand positioning.

[0052] Domain set D: The allowed value range of each variable. For example, the discount range [dmin, dmax] is determined by brand tone and product gross profit constraints. Channel priority is taken from the permutation or weight vector of {APPPush, SMS, WeChat, in-site message}.

[0053] The constraint set C includes hard constraints Ch and soft penalty functions Cs, which are used to filter legal solutions and guide the solution bias. Among them, hard constraints Ch include prohibiting further pushes when marketing fatigue reaches a preset threshold, prohibiting the use of unsubscribed or blocked channels, ensuring that the discount is not lower than the brand's minimum allowed discount, and requiring recommended products to be available and in sufficient stock. If any hard constraint is not met, the candidate solution is discarded.

[0054] The soft objective function is defined as a comprehensive expected return weighted by the priority of business objectives, including at least the estimated positive feedback response rate, negative feedback risk penalty, expected profit, and vector similarity matching value with typical user cluster centers. First, users of inappropriate size are eliminated based on marketing fatigue and unsubscription flags. Then, the remaining users are mapped to the most matching typical user clusters based on the vector similarity matching value. For each cluster, candidate strategy parameter combinations are enumerated or sampled within the value range defined by hard constraints. Combinations that do not meet the hard constraints are filtered out, and the soft objective function value is calculated for the remaining feasible solutions. The candidate combination that maximizes the soft objective function value is selected as the marketing strategy parameter for that cluster and output.

[0055] The online version of the constraint satisfaction solver, namely the online lightweight student model, accepts the user's state vector input, internally executes or approximates the above constraint satisfaction solution process, and outputs marketing strategy parameters that satisfy hard constraints and are better under the soft objective function. Online reinforcement learning, counterfactual exploration, or knowledge distillation are all used to adjust the fitting parameters of the online lightweight student model for soft penalty weights, discount interval relaxation factors, and channel priority bias.

[0056] In this embodiment, the business objective priority refers to the set of objective weights pre-configured by marketing operations personnel when creating or editing a marketing campaign. This set indicates the emphasis of the campaign among various business metrics and includes at least the normalized weights of several items among conversion rate, average order value, marketing cost control, brand exposure, and user retention rate. The business objective priority is stored in the marketing plan configuration table.

[0057] In this embodiment, the generative large model refers to a pre-trained large language model (LLM) with text generation capabilities, such as the GPT series, ChatGLM, ERNIE, or Hunyuan based on the Transformer architecture. It is configured to receive marketing prompt text obtained through structured encoding and, in conjunction with the brand tone parameters (tone, style, prohibited words) defined in the system prompt, generate natural language marketing content that conforms to the characteristics of the target user group and the marketing objectives, including at least product recommendation descriptions, promotional copy, and outreach messages.

[0058] The structured strategy parameters that satisfy the solver's output constraints are assembled into a text object (Prompt) according to a predefined prompt template, consisting of "natural language instructions + structured constraint blocks + output format requirements". Based on the user segmentation, product, discount, channel, and brand tone constraints given in the Prompt, the large language model generates marketing plan text or structured marketing content that meets the requirements. The encoding process is not limited to natural language concatenation; ChatML, XML markup, or functions can also be used. The calling format encapsulates the same semantic information.

[0059] In some optional embodiments, for step S106, the system automatically selects different learning methods based on the user's level of response to the marketing content, i.e., the feedback density. If the feedback density is higher than a first preset threshold, it indicates active response, and online reinforcement learning is used to update the parameters of the solver to satisfy the constraints.

[0060] If the feedback density is not higher than the first preset threshold and not lower than the second preset threshold, it indicates that the response is average. At this time, the user vector similarity is calculated based on the core behavior representation field to perform similar user clustering, and the counterfactual strategy is executed to update the parameters of the constraint-satisfying solver. The first preset threshold is greater than the second preset threshold.

[0061] If the feedback density is lower than the second preset threshold, it means that there are very few responses. At this time, the offline retraining process is triggered to reconstruct the offline teacher model, and the knowledge of the offline teacher model is transferred to the online lightweight student model through knowledge distillation.

[0062] The parameters of the offline teacher model are frozen. Taking the user state sampling set as input, the parameters of the online lightweight student model are updated with the goal of minimizing the difference between the output distribution of the online lightweight student model and the output distribution of the offline teacher model. Each user state vector in the user state sampling set is obtained by concatenating or encoding the core behavioral representation field, vector similarity matching value, life cycle stage field, recent positive feedback response rate, recent negative feedback response rate, and marketing fatigue field encapsulated by the dynamic user profile of the corresponding user.

[0063] Based on the same inventive concept, embodiments of the present invention provide an AI intelligent agent construction system for marketing scenarios. Figure 2 This is a structural block diagram of an AI intelligent agent construction system 200 for marketing scenarios, provided as an embodiment of the present invention. Figure 2 As shown, the AI ​​intelligent agent construction system 200 for marketing scenarios mainly includes: The multi-source data anonymization module 201 is used to anonymize the multi-source raw data collected from various marketing touchpoints to generate an anonymized dataset. The multi-source raw data includes user behavior data streams, user transaction data streams, and device environment streams. The anonymized dataset includes anonymized behavior datasets, anonymized transaction datasets, and anonymized device environment metadata. The global user identifier generation module 202 is used to calculate the similarity of behavior patterns based on the de-identified behavior dataset and the de-identified device environment metadata, construct a cross-device user identifier association graph based on the behavior pattern similarity, generate multiple global user identifiers, and bind the multiple global user identifiers as association indexes to the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata respectively. The dynamic user profile generation module 203 is used to generate a corresponding dynamic user profile for each global user identifier, using the global user identifier as the primary key; wherein, the dynamic user profile encapsulates vector similarity matching values; The constraint satisfaction solution module 204 is used to input the vector similarity matching value and the preset business objective priority into the constraint satisfaction solver to obtain the strategy parameters; The marketing plan generation module 205 is used to encode strategy parameters into semi-structured prompt text, input the semi-structured prompt text into the generative big model, and output the marketing plan so that the multi-channel reach gateway can execute the marketing plan to reach users. The feedback parameter tuning module 206 is used to collect positive and negative feedback data from users regarding marketing interventions, calculate the feedback density, and adaptively select an update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver. The feedback density is the ratio of the sum of the number of times users generate positive and negative feedback within a preset first sliding time window to the total number of marketing interventions issued during the same period.

[0064] In some optional embodiments, the multi-source data desensitization module 201 is specifically used to identify data to be classified in user behavior data and select a desensitization strategy for the data to be classified based on data distribution characteristics; wherein, the data to be classified includes direct identification information, indirect identification information, and behavior frequency statistics; if the data to be classified is direct identification information, a first desensitization strategy is selected, which is to perform irreversible hash replacement or encrypted erasure processing on the direct identification information; if the data to be classified is indirect identification information, a second desensitization strategy is selected, which includes automatically binning continuous values ​​into discrete intervals according to data distribution, truncating timestamps to hourly or daily levels according to business scenarios, and generalizing geographical locations to city or provincial levels according to population density; if the data to be classified is behavior frequency statistics, a third desensitization strategy is selected, which is to inject differential privacy noise into the behavior frequency statistics.

[0065] In some optional embodiments, the global user identifier generation module 202 is specifically used to read multiple records that have not been assigned a global user identifier from the de-identified behavior dataset, extract the behavior metadata of each record, and aggregate them at the granularity of session identifier or device fingerprint hash to obtain a behavior pattern vector; wherein, the behavior metadata includes session identifier, the identifier of the interacted object, the behavior action type, and the binned page dwell time; Calculate the behavioral pattern similarity between any two behavioral pattern vectors; Each session identifier or device fingerprint hash is used as a node in the cross-device user identifier association graph. An undirected edge is added between two nodes whose connection behavior pattern similarity is higher than a preset similarity threshold to obtain the cross-device user identifier association graph. Perform connectivity component analysis or community discovery algorithms on the cross-device user identifier association graph, merge all nodes in the same connectivity component or the same community into the same logical user, assign a unique global user identifier to each logical user, and write back the global user identifier to the index field of the corresponding record in the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata. For any node in the cross-device user identifier association graph that has not been assigned a global user identifier, obtain the set of candidate global user identifiers for all nodes connected to that node. Calculate the initial attribution probability of that node for each candidate global user identifier based on the edge weights. Then, perform a decay correction on each initial attribution probability based on the shared device factor to obtain the corrected attribution probability. Here, the shared device factor is the number of different global user identifiers bound to the device fingerprint hash of the physical terminal to which the node belongs; and the edge weight is the behavioral pattern similarity between the node connected to it and the node. If the candidate global user identifier set is empty or the maximum value of the corrected attribution probability is lower than the preset merging threshold, a new global user identifier will be assigned to the node; otherwise, the candidate global user identifier corresponding to the maximum value of the corrected attribution probability will be used as the global user identifier of the node.

[0066] In some optional embodiments, the dynamic user profile also encapsulates core behavioral representation fields, lifecycle stage fields, recent positive feedback response rate, recent negative feedback response rate, and marketing fatigue fields; the dynamic user profile generation module 203 includes: The first field generation module is used to obtain the long-term attribute field, the short-term interest field, and the real-time intent field respectively; to encode and normalize the long-term behavior representation vector in the long-term attribute field, the short-term interest vector in the short-term interest field, and the real-time behavior embedding vector in the real-time intent field respectively; and to perform weighted fusion of the normalized vectors according to the preset weight coefficients to generate a fixed-dimensional behavior pattern vector as the core behavior representation field. The second field generation module is used to read the target user's historical transaction data from the anonymized transaction dataset. Based on the historical transaction data and preset time window rules, it determines the customer relationship stage of the target user and marks the customer relationship stage as the lifecycle stage field. The historical transaction data includes registration time, first transaction time, last transaction time and cumulative number of transactions. The third field generation module is used to calculate the similarity between the core behavior representation field and the preset typical user group center vector to obtain the vector similarity matching value. The fourth field generation module is used to calculate the proportion of positive feedback generated by users within the preset second sliding time window to the total number of effective marketing campaigns delivered during the same period, which is used as the recent positive feedback response rate. The fifth field generation module is used to calculate the proportion of the number of times users generate negative feedback within the preset third sliding time window to the total number of effective marketing campaigns during the same period, as the recent negative feedback response rate; The sixth field generation module is used to calculate marketing fatigue based on whether the user has negative feedback behavior within the preset fourth sliding time window, the preset upper limit of marketing intervention allowed per unit time, and the total number of effective marketing campaigns delivered during the same period.

[0067] Further, the first field generation module is specifically used to configure a lightweight sequence model, extract the behavior sequence of the current session from the desensitized behavior dataset, encode and generate a real-time behavior embedding vector as a real-time intent field; and / or; configure a nearline computing engine, call a time decay function to extract records in the desensitized behavior dataset that match the global user identifier within the most recent preset time period; group and aggregate the matched records based on the category identifier or inventory unit identifier, and assign different behavior weights to click behavior and add-to-cart behavior according to the behavior type, and then perform weighted summation to obtain the interest score for each category or inventory unit; correct the interest score based on the binned page dwell time; generate a short-term interest vector as a short-term interest field based on the corrected interest score for each category or inventory unit; and / or; configure an offline computing engine, call a value calculation model, take the core behavior representation field and the historical transaction aggregation features in the desensitized transaction dataset that match the global user identifier as input data, and output value stratification labels and lifecycle stage labels as long-term attribute fields; wherein, the historical transaction aggregation features include consumption frequency, consumption amount and recent purchase interval.

[0068] In some optional embodiments, the feedback parameter tuning module 206 includes: The first parameter tuning module is used to update the parameters of the constraint-satisfying solver by using online reinforcement learning if the feedback density is higher than the first preset threshold. The second parameter tuning module is used to calculate user vector similarity based on the core behavior representation fields encapsulated in the dynamic user profile to perform similar user clustering, and to perform counterfactual strategy exploration to update the parameters of the constraint-satisfying solver if the feedback density is not higher than the first preset threshold and not lower than the second preset threshold; wherein, the first preset threshold is greater than the second preset threshold. The third parameter tuning module is used to trigger the offline retraining process to reconstruct the offline teacher model if the feedback density is lower than the second preset threshold, and to transfer the knowledge of the offline teacher model to the online lightweight student model through knowledge distillation; wherein, the online lightweight student model is an online running version of the constraint satisfaction solver.

[0069] Furthermore, the third parameter tuning module is specifically used to freeze the parameters of the offline teacher model. Taking the user state sampling set as input, it updates the parameters of the online lightweight student model with the goal of minimizing the difference between the output distribution of the online lightweight student model and the output distribution of the offline teacher model. Each user state vector in the user state sampling set is obtained by concatenating or encoding the core behavioral representation fields, vector similarity matching values, life cycle stage fields, recent positive feedback response rates, recent negative feedback response rates, and marketing fatigue fields encapsulated by the dynamic user profile of the corresponding user.

[0070] The functional modules in the embodiments of this invention can be integrated together to form an independent unit, such as integrated into a processing unit, or each module can exist physically separately, or two or more modules can be integrated to form an independent unit. The integrated unit can be implemented in hardware or as a software functional unit. If the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0071] Various variations and specific examples of the methods provided in the embodiments of the present invention are also applicable to the AI ​​intelligent agent construction system for marketing scenarios provided in this embodiment. Through the foregoing detailed description of the AI ​​intelligent agent construction method for marketing scenarios, those skilled in the art can clearly understand the implementation method of the AI ​​intelligent agent construction system for marketing scenarios in this embodiment. For the sake of brevity, it will not be described in detail here.

[0072] Figure 3 This is a structural block diagram of an electronic device 300 provided in an embodiment of the present invention. Figure 3 As shown, the electronic device 300 includes a memory 301, a processor 302, and a communication bus 303; the memory 301 and the processor 302 are connected through the communication bus 303.

[0073] The memory 301 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 301 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function, and instructions for implementing the AI ​​agent construction method for marketing scenarios provided in the above embodiments, etc. The data storage area may store data involved in the AI ​​agent construction method for marketing scenarios provided in the above embodiments, etc.

[0074] Processor 302 may include one or more processing cores. Processor 302 executes instructions, programs, code sets, or instruction sets stored in memory 301, and calls data stored in memory 301 to perform various functions and process data as described in this application. Processor 302 may be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), controller, microcontroller, and microprocessor. It is understood that for different devices, the electronic devices used to implement the functions of processor 302 may also be other types, and this embodiment of the invention does not specifically limit the specific devices used.

[0075] The communication bus 303 may include a path for transmitting information between the aforementioned components. The communication bus 303 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 303 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 The symbol is represented by only one double arrow, but this does not indicate that there is only one bus or one type of bus. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0076] This invention also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described in the above embodiments, which is an AI agent construction method for marketing scenarios.

[0077] In this embodiment, the computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. The computer-readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof. Specifically, the computer-readable storage medium can be a portable computer disk, a hard disk, a USB flash drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), staging random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, optical disk, magnetic disk, mechanical encoding device, or any combination thereof.

[0078] The computer program in this embodiment includes functions for executing... Figure 1 The program code for the method shown may include instructions corresponding to the execution of the method steps provided in the above embodiments. The computer program may be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network (e.g., the Internet, local area network, wide area network, and / or wireless network) to an external computer or external storage device. The computer program may be executed entirely on the user's computer as a standalone software package.

[0079] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0080] Additionally, it should be understood that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for constructing an AI agent for marketing scenarios, characterized in that, include: The collected raw data from multiple marketing touchpoints is anonymized to generate an anonymized dataset; wherein, the raw data from multiple sources includes user behavior data stream, user transaction data stream, and device environment data stream; the anonymized dataset includes anonymized behavior dataset, anonymized transaction dataset, and anonymized device environment metadata; The behavior pattern similarity is calculated based on the de-identified behavior dataset and the de-identified device environment metadata. A cross-device user identifier association graph is constructed based on the behavior pattern similarity, multiple global user identifiers are generated, and the multiple global user identifiers are used as association indexes and bound to the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata respectively. Using the global user identifier as the primary key, a corresponding dynamic user profile is generated for each global user identifier; wherein, the dynamic user profile encapsulates a vector similarity matching value; The strategy parameters are obtained by solving the solver by satisfying the vector similarity matching value and the preset business objective priority input constraint. The strategy parameters are encoded into semi-structured prompt text, the semi-structured prompt text is input into a generative large model, and a marketing plan is output so that the multi-channel reach gateway can execute the marketing plan to reach users. Collect positive and negative feedback data from users regarding marketing interventions, calculate the feedback density, and adaptively select an update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver; wherein, the feedback density is the ratio of the sum of the number of times users generate positive and negative feedback within a preset first sliding time window to the total number of marketing interventions issued during the same period.

2. The method as described in claim 1, characterized in that, The process of de-identifying the multi-source raw data collected from various marketing touchpoints to generate a de-identified dataset includes: Identify the data to be classified in the user behavior data, and select a de-identification strategy for the data to be classified based on the data distribution characteristics; wherein, the data to be classified includes direct identification information, indirect identification information, and behavior frequency statistics. If the data to be classified is the direct identification information, then the first desensitization strategy is selected, which is to perform irreversible hash replacement or encrypted erasure processing on the direct identification information; If the data to be classified is the indirect identification information, then the second desensitization strategy is selected. The second desensitization strategy includes automatically binning continuous values ​​into discrete intervals according to data distribution, truncating timestamps to hourly or daily levels according to business scenarios, and generalizing geographical locations to city or provincial levels according to population density. If the data to be classified is the behavior frequency statistics, then the third desensitization strategy is selected, which is to inject differential privacy noise into the behavior frequency statistics.

3. The method as described in claim 1 or 2, characterized in that, The process involves calculating behavioral pattern similarity based on the de-identified behavioral dataset, constructing a cross-device user identifier association graph based on the behavioral pattern similarity, and generating multiple global user identifiers, including: Multiple records without assigned global user identifiers are read from the de-identified behavior dataset. Behavioral metadata for each record is extracted and aggregated at the granularity of session identifier or device fingerprint hash to obtain a behavior pattern vector. The behavioral metadata includes the session identifier, the identifier of the interacted object, the type of behavior action, and the binned page dwell time. Calculate the behavioral pattern similarity between any two behavioral pattern vectors; Each session identifier or device fingerprint hash is used as a node in the cross-device user identifier association graph. An undirected edge is added to connect two nodes whose behavior pattern similarity is higher than a preset similarity threshold to obtain the cross-device user identifier association graph. Perform connected component analysis or community discovery algorithm on the cross-device user identifier association graph, merge all nodes in the same connected component or the same community into the same logical user, and assign a unique global user identifier to each logical user. At the same time, write the global user identifier back to the index field of the corresponding record in the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata. For any node in the cross-device user identifier association graph that has not been assigned a global user identifier, obtain the set of candidate global user identifiers for all nodes connected to that node, calculate the initial attribution probability of that node for each candidate global user identifier based on the edge weights, and perform attenuation correction on each of the initial attribution probabilities based on the shared device factor to obtain the corrected attribution probability; wherein, the shared device factor is the number of different global user identifiers bound to the device fingerprint hash of the physical terminal to which the node belongs; the edge weight is the behavioral pattern similarity between the node connected to it and the node; If the candidate global user identifier set is empty or the maximum value of the corrected attribution probability is lower than the preset merging threshold, a new global user identifier is assigned to the node; otherwise, the candidate global user identifier corresponding to the maximum value of the corrected attribution probability is used as the global user identifier of the node.

4. The method as described in claim 1 or 2, characterized in that, The dynamic user profile also encapsulates core behavioral representation fields, lifecycle stage fields, recent positive feedback response rate, recent negative feedback response rate, and marketing fatigue fields; the step of generating a corresponding dynamic user profile for each global user identifier as the primary key includes: Long-term attribute field, short-term interest field and real-time intent field are obtained respectively; the long-term behavior representation vector in the long-term attribute field, the short-term interest vector in the short-term interest field and the real-time behavior embedding vector in the real-time intent field are encoded and normalized respectively; the normalized vectors are weighted and fused according to preset weight coefficients to generate a fixed-dimensional behavior pattern vector as the core behavior representation field. Historical transaction data of the target user is read from the anonymized transaction dataset. Based on the historical transaction data and preset time window rules, the customer relationship stage of the target user is determined and the customer relationship stage is marked as the lifecycle stage field. The historical transaction data includes registration time, first transaction time, last transaction time and cumulative number of transactions. The similarity between the core behavior representation field and the preset typical user group center vector is calculated to obtain the vector similarity matching value; The proportion of positive feedback generated by users within a preset second sliding time window to the total number of effective marketing campaigns delivered during the same period is calculated as the recent positive feedback response rate. The proportion of the number of times users generate negative feedback within a preset third sliding time window to the total number of effective marketing campaigns delivered during the same period is calculated as the recent negative feedback response rate. The marketing fatigue level is calculated based on whether the user exhibits negative feedback behavior within the preset fourth sliding time window, the preset upper limit for marketing intervention per unit time, and the total number of effective marketing campaigns delivered during the same period.

5. The method as described in claim 4, characterized in that, The real-time intent field is obtained using the following method: Configure a lightweight sequence model to extract the behavior sequence of the current session from the de-identified behavior dataset, encode it to generate a real-time behavior embedding vector as the real-time intent field; and / or; The short-term interest field is obtained using the following method: Configure a nearline computing engine, call the time decay function to extract records from the de-identified behavior dataset that match the global user identifier within the most recent preset time period; group and aggregate the matching records based on category identifier or inventory unit identifier, and assign different behavior weights to click behavior and add-to-cart behavior according to behavior type, and then perform weighted summation to obtain the interest score for each category or inventory unit; correct the interest score based on the page dwell time after binning; generate a short-term interest vector as the short-term interest field based on the corrected interest score for each category or inventory unit; and / or; The long-term attribute field is obtained using the following method: Configure an offline computing engine, call the value computing model, and use the core behavior representation field and the historical transaction aggregation features in the anonymized transaction dataset that match the global user identifier as input data. Output the value stratification label and the life cycle stage label as the long-term attribute field; wherein, the historical transaction aggregation features include consumption frequency, consumption amount and recent purchase interval.

6. The method as described in claim 4, characterized in that, The adaptive selection and update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver includes: If the feedback density is higher than the first preset threshold, then online reinforcement learning is used to update the parameters of the constraint satisfaction solver. If the feedback density is not higher than the first preset threshold and not lower than the second preset threshold, then the user vector similarity is calculated based on the core behavior representation field encapsulated by the dynamic user profile to perform similar user clustering, and counterfactual strategy exploration is performed to update the parameters of the constraint satisfaction solver; wherein, the first preset threshold is greater than the second preset threshold; If the feedback density is lower than the second preset threshold, an offline retraining process is triggered to reconstruct the offline teacher model, and the knowledge of the offline teacher model is transferred to the online lightweight student model through knowledge distillation; wherein, the online lightweight student model is the online running version of the constraint satisfaction solver.

7. The method as described in claim 6, characterized in that, The process of transferring knowledge from the offline teacher model to the online lightweight student model through knowledge distillation includes: The parameters of the offline teacher model are frozen. Taking the user state sampling set as input, the parameters of the online lightweight student model are updated with the goal of minimizing the difference between the output distribution of the online lightweight student model and the output distribution of the offline teacher model. Each user state vector in the user state sampling set is obtained by concatenating or encoding the core behavior representation field, vector similarity matching value, life cycle stage field, recent positive feedback response rate, recent negative feedback response rate, and marketing fatigue field encapsulated by the dynamic user profile of the corresponding user.

8. An AI intelligent agent construction system for marketing scenarios, characterized in that, include: The multi-source data anonymization module is used to anonymize the raw data collected from various marketing touchpoints from multiple sources to generate an anonymized dataset; wherein, the raw data from multiple sources includes user behavior data streams, user transaction data streams, and device environment streams; the anonymized dataset includes an anonymized behavior dataset, an anonymized transaction dataset, and anonymized device environment metadata; The global user identifier generation module is used to calculate the similarity of behavior patterns based on the de-identified behavior dataset and the de-identified device environment metadata, construct a cross-device user identifier association graph based on the behavior pattern similarity, generate multiple global user identifiers, and bind the multiple global user identifiers as association indexes to the de-identified behavior dataset, the de-identified transaction dataset, and the de-identified device environment metadata respectively; The dynamic user profile generation module is used to generate a corresponding dynamic user profile for each global user identifier, using the global user identifier as the primary key; wherein, the dynamic user profile encapsulates a vector similarity matching value; The constraint satisfaction solution module is used to input the vector similarity matching value and the preset business objective priority into the constraint satisfaction solver to obtain the strategy parameters; The marketing plan generation module is used to encode the strategy parameters into semi-structured prompt text, input the semi-structured prompt text into the generative big model, and output the marketing plan so that the multi-channel reach gateway can execute the reach of the marketing plan; The feedback parameter tuning module is used to collect positive and negative feedback data from users regarding marketing interventions, calculate the feedback density, and adaptively select an update strategy based on the feedback density to adjust the parameters of the constraint-satisfying solver. The feedback density is the ratio of the sum of the number of times users generate positive and negative feedback within a preset first sliding time window to the total number of marketing interventions issued during the same period.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the AI ​​agent construction method for marketing scenarios as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the AI ​​agent construction method for marketing scenarios as described in any one of claims 1 to 7.

Images